Research map / Detection and segmentation / CNN features
DynamicDet: A Unified Dynamic Architecture for Object Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use adaptive routing so only 50% of images take the second detector | Dy-YOLOv7-W6 · COCO test-dev | AP: 56.0 -> 56.1 | Helped |
| Use adaptive routing so only 50% of images take the second detector | Dy-YOLOv7-W6 · COCO test-dev | FPS: 52 -> 58 | Helped |
| Use adaptive routing so all images take the second detector | Dy-YOLOv7-W6 · COCO test-dev | FPS: 33 -> 46 | Helped |
| Use the proposed loss-offset router training instead of random selection (ablation) | Dy-Mask R-CNN Swin-T · COCO test-dev, 84.5 ms latency | AP_box: +0.9% AP | Helped |
| Use the proposed loss-offset router training instead of AP-based router training (ablation) | Dy-Mask R-CNN Swin-T · COCO test-dev, 84.5 ms latency | AP_box: +0.7% AP | Helped |
| Use a validation set of 1k rather than the test set for the 50% routing threshold (ablation) | Dy-Mask R-CNN Swin-T · COCO | threshold deviation: 0.11 at 1k | Hurt |
| Increase the proportion of images routed through the second detector (ablation) | Dy-YOLOv7-W6 · COCO test-dev | AP: 54.7 -> 56.8 | Helped |
| Increase the proportion of images routed through the second detector (ablation) | Dy-YOLOv7-W6 · COCO test-dev | FPS: 78 -> 46 | Hurt |
| Use dynamic inference instead of changing input resolution | Dy-YOLOv7-W6 · COCO test-dev | AP: 52.2 -> 55.2 | Helped |
| Extend the dynamic architecture to a two-stage detector | Faster R-CNN ResNet50 · COCO minival | AP_box: 39.4 -> 40.4 | Helped |
| Extend the dynamic architecture to a transformer-backbone detector | Mask R-CNN Swin-T · COCO minival | AP_box: 48.2 -> 49.9 | Helped |
Closest papers by meaning
- Dynamic Refinement Network for Oriented and Densely Packed Object Detection (CVPR 2020 · similarity 0.89)
- EfficientDet: Scalable and Efficient Object Detection (CVPR 2020 · similarity 0.88)
- SpeedDETR: Speed-aware Transformers for End-to-end Object Detection (ICML 2023 · similarity 0.87)
- Test-Time Adaptive Object Detection with Foundation Model (NeurIPS 2025 · similarity 0.87)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.85)
- DCSF-KD: Dynamic Channel-wise Spatial Feature Knowledge Distillation for Object Detection (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.